The Laya engine is presented for making structured decisions

Developers introduced Laya — a specialized engine for processing typed queries, which eliminates the text generation stage. The system is focused on working with arbitrary data states and providing strictly structured answers.

The key feature of the architecture is a non-autoregressive approach, allowing for query evaluation in a single direct pass. The processing time for one query is around 35 ms, which is significantly faster than standard LLMs that require sequential token generation.

In addition to forming answers, the tool automatically calculates calibration confidence scores. This allows the system to signal the degree of reliability of the decision made in real-time. The project is published in open access on GitHub under a license that allows integration into high-load systems.

The absence of a generative stage makes the solution suitable for tasks where predictability of output and low latency are critical. The technology is focused on automating business logic and analytical processes that require instant response to system state changes without spending computational resources on building natural language.